A chat with a friend in robotics made me see AI training differently

For years, I thought training an AI was like teaching a single, very smart dog a new trick. I was helping my friend, who works on robotic arms, sort parts for a project last month. He explained his team uses a method where the AI learns from thousands of simulated, failed attempts before it ever touches a real screw. The change for me was understanding it's less about direct instruction and more about creating a whole world of practice for it to explore on its own. Has anyone else had their view on how AI learns shift from a specific conversation?

3 comments

3 Comments

finleyh892mo ago

Got a different take from my own tinkering. That simulation world is still built by people giving it very specific rules and goals. Feels less like free exploration and more like a really complex maze we built for it, doesn't it?

the_charlie2mo ago

Look at it from the AI's point of view though, @finleyh89. Even if the maze is built by us, the way it moves through it can still be its own. I mean, we're born into a world with physics and social rules we didn't pick, but we still find new paths. Maybe the test isn't if the maze is there, but what it does when it hits a dead end we never saw coming.

gracewebb2mo ago

Wait, you actually tinker with AI, @finleyh89?